Video Transcript — Whisper-Powered, Free

Whisper-1 is one of the best general-purpose speech-to-text models available. For clear English speech in a quiet environment, word error rate is around 3–5%. Accents, background noise, multiple speakers, and technical jargon can drop accuracy to 10–20%. The transcript includes speaker-attributed segments with start and end timestamps.

Whisper supports 99 languages including English, Spanish, French, German, Portuguese, Italian, Dutch, Russian, Chinese (Simplified and Traditional), Japanese, Korean, Arabic, Hindi, and many more. The model auto-detects the language; you can also pass a language hint if auto-detection picks the wrong one.

A transcript turns a video into text you can search, copy, and quote. Common uses: Search inside a long recording. A 90-minute meeting is hard to scrub through. A searchable transcript turns it into Ctrl+F territory. Subtitles for a video. The .srt output drops into VLC, YouTube, Handbrake, or any video editor. Hardcoded or softcoded subtitles in two clicks. A written record of an interview. A journalist or researcher can transcribe a 60-minute interview in three minutes, then spend the time saved on the actual analysis. Note-taking from a talk. A conference talk, a podcast, a YouTube video you downloaded — get the text and quote the parts that matter.

We use OpenAI's Whisper-1 model via the same endpoint that powers Pullsy Record's automatic transcription. The pipeline: 1. Browser sends the uploaded file to `record.pullsy.online/api/transcribe-upload` (multipart POST). 2. If the file is under 24 MB, it goes to Whisper directly. If over 24 MB, we extract audio as 16 kHz mono MP3 first (matching the existing `/api/transcribe/:id` pipeline). 3. Whisper returns verbose JSON with `segments` (each with `start`, `end`, `text`) and a `text` field with the full transcript. 4. We return `{ segments, fullText, durationSec, language }` to the browser. 5. The browser renders the segments with timestamps, plus a 'Copy text' button and downloads for .txt and .srt. We do not store the uploaded file. The audio extraction step (for large files) goes through the same Polish tunnel that Pullsy Record uses for transcribing its own recordings.

Your file is uploaded to our transcription endpoint (a Cloudflare Worker) which forwards it to OpenAI's Whisper-1 API for transcription. We do not store the file on our servers. OpenAI retains Whisper API data for 30 days for abuse monitoring, then deletes it (per OpenAI's API data usage policy). If your file contains sensitive content and you cannot send it to OpenAI, use the Pullsy Record flow instead — record directly in Pullsy and the audio never leaves Cloudflare R2.

Yes — drop an MP3, WAV, or M4A file. Whisper handles audio-only inputs natively. The 50 MB cap applies to the input file; for a 1-hour mono MP3 at 64 kbps, the file is ~28 MB, well within the limit.

Whisper-1 does not produce speaker diarization. The transcript is one continuous text with timestamps. If you need speaker labels, you need a separate diarization model (like pyannote.audio) — we do not have that tool today.

Whisper auto-detects the language from the first 30 seconds. For better accuracy on a non-English file, you can specify the language with a hint — the API supports a `language` parameter. Our current implementation uses auto-detection; a language picker is on the roadmap.